Since 2 August, if your event app has a chatbot, you have been legally obliged to tell people it is one.
Four days before that, the rule that would have made you prove what that chatbot actually did moved to December 2027.
Almost nobody in events noticed either thing, as far as I've seen on LinkedIn and online. I think it's because both arrived inside a document called the “Digital Omnibus on AI,” which is definitely one of the least readable phrases anyone has produced this year.
So, here's the whole story, in order, in simple words and terms. Including the part I got wrong, which turned out to be the most interesting part.
I’m not a lawyer and this is most definitely not legal advice. It's a systems argument with dates attached.
First, the four words you need
The EU AI Act is one law covering artificial intelligence across every industry. It doesn't care whether you're a bank or a congress organiser. It cares what the AI does and who it does it to.
A provider builds the thing and puts it on the market. Your platform. Your app vendor. Us, when we build something.
A deployer uses it in a professional context. That's the organiser. That's you.
High-risk doesn't mean dangerous in the ordinary sense. It's a specific list in a specific annex. Annex III covers:
- biometric identification
- recruitment and worker management
- education admission and assessment
- access to essential services
- law enforcement and border control
If your AI does one of those, a long list of obligations attaches. Technical documentation. Event logging. Human oversight. Record-keeping. Post-market monitoring.
Transparency is the lighter tier, and it lives in Article 50.
It isn't about proving anything.
It's about saying something.
Tell people it's an AI. Mark AI-generated content as AI-generated. Tell people when you're scanning their faces or reading the room.
Transparency is the label. Documentation is the logbook.
One of those is a wet floor sign. The other is a mop.
Please hold on to that, because the whole story is about which one turned up.
The timeline
1 August 2024. The AI Act enters into force. The obligations don't all start at once. They roll out over four years, which is how you get a law that is simultaneously two years old and not yet properly switched on.
February and August 2025. The outright bans start. Social scoring, certain manipulation and some emotion recognition in workplaces and schools. Rules for general-purpose AI models follow six months later, and the EU AI Office opens for business. A busy period all round.
19 November 2025. The Commission admits, in public and in writing, that implementation is off track and tables the Digital Omnibus on AI. The harmonised standards weren't written. The notified bodies weren't designated. The infrastructure the high-risk regime needed in order to function didn't exist.
Stay with me here; I know this is boring stuff. We're nearly through it.
28 April 2026. Talks collapse.
6 May 2026. Talks uncollapse. Provisional agreement, confirmed by member states on 13 May.
24 July 2026. Regulation (EU) 2026/1744 is published in the Official Journal, a title that really sings. It is the first substantive amendment the AI Act has ever had.
27 July 2026. It enters into force.
2 August 2026. Sunday. Two things were supposed to happen. One of them did.
What landed on Sunday
Article 50. The label. The wet floor sign.
If you're a provider and your system talks directly to a person, it must be built so that person knows they're talking to an AI. Unless it's blindingly obvious to a reasonably well-informed person, in which case you can leave it. I still wouldn't though.
If you're a provider and your system generates synthetic audio, images, video or text, the output must be machine-readable as artificially generated. Anything already on the market before Sunday gets until 2 December 2026.
If you're a deployer running emotion recognition or biometric categorisation, you must tell the people it's pointed at.
And if you're publishing deepfakes or AI-generated text on matters of public interest, you must disclose it.
That's Sunday.
Say the thing.
Mark the thing.
Tell the room.
What did not
The Annex III high-risk obligations were due the same day. They moved to 2 December 2027. AI embedded in regulated products, under Annex I, moved to 2 August 2028.
Those are the obligations about documentation, logging and oversight. Being able to reconstruct what a system did and why.
The duty to say arrived on time.
The duty to prove moved sixteen months to the right.
There was also a lovely small detail almost nobody has picked up as far as I've seen. The duty to make sure your staff understood AI was softened into a duty to support your staff in understanding AI.
Read that twice. I did.
Now for the part I got wrong. Twice.
I had written that facial recognition at check-in was heading for high-risk classification in December 2027 and that organisers using it should definitely get that date in their diaries.
Then I read Annex III point 1(a) properly and decided I'd got it wrong.
It makes remote biometric identification high-risk, and then, in the same breath, excludes biometric verification whose sole purpose is confirming that someone is who they claim to be. That sounded exactly like a check-in kiosk to me. One face, one record, are you who you say you are. Carved out by name.
So, I went along and had my homework marked.
Panos Moutafis, Ph.D. has a doctorate in this and has been building facial analysis at Zenus AI (Zenus, Inc.) for the events industry for the best part of a decade. We've known each other since the early days of it, back when the check-in kiosks we were using were stacked white boxes with a NUC, screen and keyboard, and I was running the technology for the African arm of a high-volume onsite check-in business. Twelve kiosks and a company that didn't survive 2020.
We started talking. I told him what I'd found and asked if this was the misunderstanding he'd mentioned before.
“There's a bigger one,” he said.
I had talked myself out of a correct answer.
“Verification is often called one-to-one. Scan my ID and take my picture. Confirm these two are the same person. Identification is often called one-to-many. Take someone's picture on a check-in kiosk and compare it against the entire database of images provided during registration. Face recognition for check-in would fall under the high-risk classification.”
Read that twice as well. I did.
The carve-out is real. It just doesn't cover what we actually do.
Verification is one to one. You hand over a passport and the machine compares your face to the photo in it. Two things. Are they the same.
Identification is one to many. You walk up to a kiosk, it has no idea who you are, and it compares your face against every photo uploaded by every registered attendee until it finds you.
The second one is what event check-in is. Which puts it back inside Annex III and puts 2 December 2027 back in the diary for anyone running it.
I found the right answer, reasoned my way out of it, and needed someone with a doctorate to pull me back. That is roughly the story of this entire regulation.
Two things the market runs together
The first is that recognition and analysis are not the same product and the Act doesn't treat them the same way.
Recognition extracts a unique signature to identify an individual.
Analysis detects that a person is present and measures the group in aggregate.
One knows you are Jane Doe. The other knows that eleven people walked past and that the room brightened when the coffee arrived.
No unique identifier, no way back to a name. And there is European Data Protection Board guidance saying that where the purpose is to distinguish one category of people from another rather than to identify anyone uniquely, you aren't in Article 9 territory under GDPR at all.
The second is the one I couldn't resolve on my own, and it turns out the Act answers it directly.
I'd found Recital 14, which says biometric data can allow for categorisation and for the recognition of emotions. That to me reads like the drafters contemplated doing this without identifying anyone, which would drag anonymous sentiment analytics into scope.
Recital 18 is the answer. The emotion recognition definition explicitly doesn’t include “the mere detection of readily apparent expressions, gestures or movements, unless they are used for identifying or inferring emotions.” A frown. A smile. A raised voice.
That last clause is doing a lot of work, and it's the whole distinction. Counting smiles as smiles is outside. Counting smiles to conclude that someone is happy is back inside.
Panos puts the human version better than the legal one.
“If you see my face, you don't know what I'm thinking. You can talk to me for five minutes and you still don't know what I'm thinking. But if I give a talk to ten thousand people and I tell a joke and everybody laughs, and then I get too confident and tell another one and nobody laughs, the group reaction, or the lack of one, is more reliable. And you are not focusing on one individual.”
As far as I've understood until now it all comes down to energy, not emotion. The regulation is aimed squarely at the other thing.
“What you don't want is somebody saying, ‘I am interviewing this person for a job, and based on their facial expression I am going to reject them.’ What the regulation is really trying to protect is the impact on individuals and their lives. For me that's a no-brainer.”
The distinction nobody in events is making
Here's the one I hadn't thought about as much as I'd like to admit, and it reframes every AI conversation you're going to have this year.
“Generative AI is stochastic. You can ask the same question and get different responses. Our machine learning models are deterministic. Same input, same output. They are also frozen. They are not using client data to self-improve. When enterprise clients onboard us, this is the first thing their compliance teams want to understand.”
Two things, both called AI, with completely different risk profiles. The first one can give you a different answer on Tuesday than it gave you on Monday. The other one can't.
It seems to me that we, the collective, have spent a year in this industry using one word for both.
The architecture matters for the same reason. Panos has stopped calling the hardware a ‘camera’, because the word implies assumptions that aren't true.
“When you say camera, what people understand is that you take a picture, or you stream video like in a video call. We don't do those things. The light hits the lens, it goes through the cable straight into the device, and we analyse it in real time without storing or transmitting images or video.”
“It is processing. But it is not capturing, storing or transmitting.”
So, most event technology was never in scope. Except the bit a lot of you bought.
Go back and read the Annex III list. Then notice what isn't on it. Roughly everything on the show floor at Event Tech Live.
A session recommender is not high-risk. A matchmaking engine is not high-risk. An AI drafting your speaker bios is not high-risk, however much the launch deck implied otherwise. Aggregate expression analysis, on the text, sits outside the emotion definition.
And then there’s facial recognition check-in. Which a great many of you have bought. Which sits squarely inside Annex III. And which most of this industry currently believes is fine.
The other genuine exceptions are narrower than people think. A careers fair where AI screens candidates is employment. A CPD programme where AI assesses or supervises is education.
Which means the honest position is this. For most of your stack, the law was never going to force an audit trail. Not because it got delayed. Because it was never pointed at you in the first place.
For the kiosk at your front door, it is pointed at you, and the date is 2 December 2027.
For everything else, the audit trail is a commercial question.
It always was.
Which is why the sixteen months is not the relief it looks like
I had originally written that the delay was probably the right call and left it there.
The standards weren't published, the notified bodies weren't designated, and regulating to a fictional timetable produces paperwork rather than safety. All still true.
But I put it to the person actually carrying the regulatory risk, and he doesn't see it as relief at all.
“People tell us the US is more relaxed and that we'll have trouble in Europe. We love doing business in Europe. GDPR and the EU AI Act are comprehensive and detailed. You know exactly what you can do and what you cannot do. Elsewhere there are regulations that just say ‘biometric information’, so broad that you don't know what is included and what is not.”
“Having a framework that explicitly says what you can and cannot do makes my life easier as a vendor, and it makes life easier for the business buying my service, because they don't have that risk sitting in their head. I wish this was adopted globally.”
That stopped me.
The vendor with the most exposure to this regulation wants more of it, sooner, everywhere.
Not because he enjoys compliance, but because ambiguity is more expensive than rules.
A detailed regime tells you where the walls are. A vague one leaves every buyer carrying unquantified risk in their head and every honest vendor competing against people who simply assert they're fine.
So, the sixteen months isn't a gift to builders. It's sixteen more months of vagueness and the people who wanted the clarity most are the ones who had already built for it.
What survives every exemption
This is the part that matters most and it came from the vendor rather than the regulator.
Analysis outside the emotion definition.
No Article 9 special category processing.
No consent required, provided you have a legitimate interest under Article 6(1)(f), which running your event properly generally is, as long as you're actually using the data rather than hoarding it like a dragon.
And then:
“Even with all of those exemptions, there are still things you have to do. You should have a data protection impact assessment. You need specific contracts saying how long you retain the data and what happens when you delete it. And you need signage. It's straightforward, but you need to do it.”
Read that list again.
A data protection impact assessment. And here’s the part organisers get wrong, which Panos was careful to correct me on. Your vendor having one does not give you one. Zenus AI holds third-party assessments and hands them over on request. You are still expected to do your own.
A contract that says what is kept and for how long and what happens when it's deleted.
A notice telling people it's happening.
That is the logbook. The mop, if you will.
Nobody is legally compelling most of it in events. He does it anyway.
On what the notice actually looks like, which is the bit you can act on this week:
“You need to put it in your privacy policy or have a link to it. And you need signage. We recommend it at registration, where people go to pick up their badges, and outside the rooms or areas where you are running the technology. That's it.”
Signage.
A QR code.
If delegates don't care, they carry on. If they do care, they were warned.
If you build
Use the time. Don't wait for it.
Retrofitting an audit trail into a system that's been running for two years is a completely different animal from designing one in. It's the difference between plumbing a new build house and plumbing a house that is already full of furniture and people.
Build the log now. Grab the mop.
Not because December 2027 is coming for you, because for most of what we build it isn't. Because your buyers will start asking before any regulator does and the first serious procurement question of 2027 is going to be some version of “show me what it did”.
The vendors who can answer that won't need to explain why it matters.
If you buy
You're a deployer. Article 50 landed on you on that Sunday when no one was watching, not in eighteen months' time.
If your event app has an assistant, check if it introduces itself as AI. If your engagement platform or device reads the room, check that the room has been told.
Small jobs this week to avoid awkward conversations later.
And if you scan faces at the door, that one isn't a small job, and the date is real.
Then the harder point.
Picture a delegate at nine at night, three days out, messaging your event app about how to get to the venue from the train station. An agent answers. From Sunday, you must tell them it's an agent.
Nothing yet requires anyone to be able to show, six weeks later, what that agent read, what it wrote back or which record it changed.
The label makes the missing logbook harder to spot, not easier. A system that announces itself feels accountable.
Announcing yourself is not the same as being able to account for yourself.
Five questions I'd ask a vendor this quarter
- Is this generative or deterministic? Can it give me a different answer tomorrow for the same question?
- Which parts of this product talk directly to an attendee and where does it say so?
- If a delegate asks in November what your assistant told them in September, can you produce it? How long do you keep it?
- What does it write, not just read? Which records can it change with no human in between?
- Can I see your data protection impact assessment and what will I need in mine?
None of those are legal questions.
They're the questions you'd ask about any system holding data you're about to turn into packed rooms, food, badge stock and transfer coaches.
What comes next
Watermarking grace ends in December.
Sandboxes in August 2027.
Annex III in December 2027.
Annex I in August 2028.
Diarise them if you'd like.
I'd hold them loosely though.
This wonderfully named Omnibus was the first amendment to this law and I’m certain it won't be the last. The pattern is now established: when the infrastructure isn't ready, the dates move.
The obligations don't vanish.
So, plan for the obligations, not the dates.
I use this technology, AI in the broadest terms, every day. All day. It's the reason I get my evenings back and none of this is an argument for slowing down.
I wouldn't have written it if it were.
I'm banging the same drum on my soapbox here again.
Anything you can't audit, you aren't managing. You are hoping.
The label is now the law.
For most of what we run in events, the logbook never was and probably never will be. If you scan faces at the door, that changes in December 2027.
For everything else, keeping one is a choice rather than a requirement.
That's the only part worth judging a vendor on.